high performance computing on graphics processing units: hgpu.org

Registrations in medical imaging and computational anatomy can be obtained using the Large Deformation Diffeomorphic Kernel Bundle Mapping (LDDKBM) framework. This provides a registration algorithm with a solid mathematical foundation while incorporating regularization of deformation at multiple scales. Because the variational formulation of LDDKBM implies a heavy computational burden in the search for optimal registrations, exploiting every possibility for faster computation will improve the usability of the algorithm. We present a parallelization strategy using the multi-scale structure and show that the parallelized method constitutes an example of how the processing power of GPUs can massively reduce the running time: after moving the computation to the GPU, we achieve a two order of magnitude speedup over a singlethreaded CPU implementation. Not only does this significantly reduce the cost of using multiple scales, it also allows the algorithm to be used on much larger datasets.